🔍 Read the full analysis: A Closer Look At Muse Spark Open’s Strength And Price on ThorstenMeyerAI.com
Get the latest gadgets delivered free with Prime
- Fast, free delivery on millions of items
- Prime Video, Amazon Music and more included
- Member-only deals all year
TL;DR
Meta released Muse Spark Open on October 15, 2026, with weights available on Hugging Face. Artificial Analysis scored it 38.9, citing leading results among open models on visual reasoning and its Omniscience measure, while testing found weaker performance on a harder terminal-agent benchmark. Its listed API prices match Meta’s closed Muse Spark 1.3, and some open alternatives cost less per index task.
Muse Spark Open is a mixture-of-experts model with 42 billion active parameters per token and 256 experts, eight of which are active for each token. Its stated context window is one million tokens. It accepts text, images, speech or audio, and video, and generates text. Meta published weights in BF16 and NVFP4 formats. Its model card estimates self-hosting requires 700 GB of GPU memory for NVFP4 or 2 TB for BF16, and says it is optimized to run on two Mac Studios with M5 Ultra using MLX.
Artificial Analysis tested the model at its default xhigh reasoning effort and recorded 38.9 on Intelligence Index v4.3.2, rounded to 39 in its reporting. The outlet ranks it tenth among open-weight models overall and says the nine models above it are from Chinese labs. Among US releases, the result is substantially above NVIDIA’s Nemotron 3.5 Super VL, which scored 26.5 in the same-day comparison cited in the source material. Meta’s closed Muse Spark 1.3 scored 45.1 at the same reasoning effort.
The strongest reported results include 24.7 on AA-Omniscience, the highest score among 167 open-weight entries in that test, and 81% on MMMU-Pro visual reasoning, the highest among 66 tested open-weight models. Artificial Analysis reported a 24% hallucination rate, the lowest among the 15 highest-scoring open models in its comparison. Results vary by task: on Terminal-Bench 4.0, it scored 8%, compared with 33% for Muse Spark 1.3, 35% for MiMo-V2.6-Pro and 42% for GLM-5.3.
Infographic 01 · Where Muse Spark Open lands
The best US open-weight model is tenth. The nine ahead are Chinese.
Artificial Analysis scores Meta’s first open Muse Spark model at 38.9. That is 12 points clear of any other US open-weight model, but behind nine Chinese models. Four of those nine are MIT-licensed, and three of them cost a third or less per task.
Top open-weight models by Intelligence Index
Higher is better. Effort variants of the same model are not listed twice. Cost = average API cost to run one index task.
License: Muse Spark Open ships under the Muse Model License: broad use and redistribution, but mandatory “Built with Muse Spark Open” attribution, a separate agreement for model-as-a-service businesses above $20M revenue, and a usage policy.
Source: Artificial Analysis, Muse Spark Open (xhigh) model page and model data, read 16 Oct 2026. Intelligence Index v4.3.2, ten evaluations. License categories as listed by Artificial Analysis. Countries: Xiaomi, Z AI, StepFun, Kimi, InclusionAI, Alibaba and DeepSeek are Chinese labs; Meta and NVIDIA are US.
Powered by Thorsten Meyer AI
Strong Reasoning, Uneven Agent Results
The test results point to a model that may suit visual reasoning and long-document work better than unattended terminal agents. Its one-million-token context window, 81% MMMU-Pro result and comparatively low reported hallucination rate form a useful combination for tasks involving mixed media or lengthy material. These are benchmark findings, though; they do not establish how the model will perform on every real-world workflow.
The weaker Terminal-Bench 4.0 score matters for teams considering it for autonomous coding or computer-use tasks. The model card describes Muse Spark Open as built for coding and agentic work, and reports results including 89.2 on MCP-Atlas, 80.4 on OSWorld Verified and 76.4% on Terminal-Bench 2.1. Artificial Analysis’ 8% score on the harder Terminal-Bench 4.0 shows why results from different benchmark versions should not be treated as interchangeable.
For buyers, the release also separates access to weights from the cost of hosted use. The weights are available at no charge, but the four third-party providers listed in the source material have a median price of $1.25 per million input tokens, $0.15 per million cached input tokens and $4.25 per million output tokens. Those rates match the cited prices for closed Muse Spark 1.3. Running the model independently also calls for substantial memory, according to Meta’s estimates.
high performance AI model hosting GPU
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Meta Opens Its Muse Model
Muse Spark Open is the first open-weight release in Meta’s flagship Muse Spark line. The source compares it with Muse Glimmer, a 30-billion-parameter model released under Apache 2.0 in August 2026 that scored 17 on the current Intelligence Index, and Llama 4 Maverick, released in April 2025, which scored 9. These comparisons show the scale of the new release within Meta’s open offerings; the scores reflect the index version cited in the source.
The broader open-weight field remains competitive. The source lists Xiaomi’s MIT-licensed MiMo-V2.6-Pro at 46.3, above Muse Spark Open’s 38.9, and cites several other models with higher scores. On the other hand, Meta’s model outscored the cited US open-weight alternatives. Artificial Analysis describes it as the leading US open-weight model, while the source qualifies the comparison by noting that permissively licensed US models are a separate category.
In the cited weighted cost per Intelligence Index task, Muse Spark Open costs $0.84. The source reports $0.13 for MiMo-V2.6-Pro, which scored 46.3, and $0.27 for DeepSeek V4.1 Flash, which scored 39.5. Those task-cost figures are distinct from per-token API prices and depend on the benchmark run’s token use. The source says 80% of Muse Spark Open’s weighted task cost came from input, including cached reads and writes.
““built for coding and agentic work””
— Meta, in the Muse Spark Open model card
AI model training GPU memory requirements
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
License and Real-World Costs
The provided source does not specify the full terms of Muse Spark Open’s license, including the conditions that distinguish it from permissive licenses such as MIT or Apache 2.0. Readers evaluating commercial use should consult the model’s published license and model card directly. The source also gives median provider prices, not a single guaranteed rate for every service or usage pattern.
Benchmark rankings and costs depend on test setup. Artificial Analysis used its default xhigh reasoning effort, and the source does not provide enough detail here to independently reconcile every score or cost across providers. The reported low hallucination rate applies to the comparison group and testing method described; it does not mean the model cannot produce incorrect answers. Its performance in specific production tasks remains unestablished by the cited results.
As an affiliate, we earn on qualifying purchases.
Checks Before Building With Muse
The model weights are available on Hugging Face, so developers can inspect the published files and model documentation, subject to the license terms. Organizations considering hosted access can compare provider-specific prices and token accounting with their expected workloads; teams considering self-hosting can weigh Meta’s memory estimates against their available hardware.
Further provider testing and independent evaluations could clarify how the model performs across real applications, especially coding and agent workflows. For now, the published evidence supports a strong result in some reasoning and vision tests, alongside a significant gap on Terminal-Bench 4.0 and hosted token prices at the level of Meta’s closed model.
open-weight AI models for visual reasoning
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Key Questions
When did Meta release Muse Spark Open?
Meta released the open-weight model on October 15, 2026, with weights posted on Hugging Face.
How did Muse Spark Open score?
Artificial Analysis reported a score of 38.9 on Intelligence Index v4.3.2 at its default xhigh reasoning effort. It rounds that result to 39 in its reporting.
Is Muse Spark Open cheaper than Meta’s closed model?
The median listed per-token API prices in the source match those for closed Muse Spark 1.3. The weights are free to access, but self-hosting requires substantial memory, according to Meta’s estimates.
What tasks appear to suit the model?
Its reported results are strong on visual reasoning and AA-Omniscience, and it has a one-million-token context window. Its 8% Terminal-Bench 4.0 score indicates weaker performance on that test of terminal-based agent work.
Is Muse Spark Open permissively licensed?
The source describes its license as having conditions but does not provide the full terms. Check Meta’s published license before deciding whether it fits a project.
Source: ThorstenMeyerAI.com
Halloween Picks
halloween
As an affiliate, we earn on qualifying purchases.
